Im working on an imbalanced data classification problem that K-mean clustering seems not to be solving well. I heard Fuzzy support vector machine can help. How do I get this algorithm to work?
In my mind, firstly you should to know the type of your data to classified. I mean it's a supervised data or unsupervised. K-mean clustering is usually for the unsupervised clustering. I think the following link could help you to understand the fuzzy support vector machine http://www.lib.kobe-u.ac.jp/repository/90000233.pdf and https://www.elen.ucl.ac.be/Proceedings/esann/esannpdf/es2002-6.pdf
Thank you. It is a supervised data and that's why i want to try other classification algorithms.
Yuo may find following books useful:
Support Vector Machines for Pattern Classification. Prof. Dr. Shigeo Abe, DOI 10.1007/978-1-84996-098-4 Springer London Dordrecht Heidelberg New York.
Introduction to Semi-Supervised Learning. Ronald J. Brachman, Yahoo! Research Thomas Dietterich,Oregon State University
Dear Adeyemi,
Before you leap into the world of fuzzy SVMs, have you tried k-means with smart seeding (i.e. many trials of initial centre estimates)? For non-well-spaced data, the k-means algorithm can get stuck in local minima. A loop with a random number generator to generate a more diverse initial seeding may just work fine on your data. I'd suggest this first as it is computationally very fast.
Cheers and good luck,
Mike
K-means is one of the worst performers under imbalanced class distributions. It may also be worthwhile if you try some methods such as svm or naive bayes before going for fuzzy solutions. It is guaranteed that they will produce better results than K-means.
Another suggestion is to use several methods for artifically balancing your training data. Instance balancing between classes usually work nicely if your data is large enough. Adding dummy instances, or picking equal number of random instances from each class in the training set usually works wonders on many cases.
- Badges
- Science topic